·5 min read
Calculating real ROI on AI automation, without the hype
Most AI ROI calculators are vendor-friendly fiction. Here's the back-of-envelope model we actually use with clients before any code gets written.
Every AI vendor has an ROI calculator on their pricing page. Almost none of them are honest. They overstate hours saved, ignore the cost of integration and change management, and quietly assume the automation works perfectly from day one.
When we sit down with a clinic, ops team, or founder, we use a simpler model. It fits on a napkin and survives contact with reality.
The four numbers that actually matter
- Volume.How many times per week does the task you want to automate happen? If the answer is "a few times a month," automation isn't your highest-leverage move.
- Loaded cost per occurrence. Not just wages — include the cost of context-switching, errors, and the opportunity cost of the task crowding out higher-value work.
- Realistic capture rate. What percentage of those occurrences will the AI actually handle end-to-end in the first 90 days? Be honest. 60% is a great result. 95% is a sales deck.
- All-in cost. Software licences plus integration time, training, and the ongoing tuning cost. The last one is the one most calculators leave out.
The formula
Annualised gain = volume × cost-per-occurrence × capture rate × 52, minus all-in cost.
That's it. If the result isn't a comfortable multiple of the all-in cost, the project isn't ready — either the volume is too low, the task is too complex for current models, or the integration cost is eating the savings. Find a different problem.
Where the model breaks down
This formula is great for cost-side automations: receptionist overflow, intake forms, recall reminders, internal triage. It's weaker for revenue-side bets, because revenue is non-linear and depends on a lot more than the automation itself. For revenue automations we run the same formula, then apply a 50% haircut and require the project to still pay back inside 90 days. Anything that can't survive that test gets de-prioritised.
Why the discipline matters
AI projects fail for two reasons: people automate the wrong thing, or they automate the right thing but never measure it. A simple, defensible model up front gives you something to measure against afterwards. Without it, you can't tell whether the automation is winning or losing — and the project quietly becomes a line item nobody wants to discuss.
If you want to run your own numbers, our ROI calculatoris the same model wrapped in a UI. It won't lie to you.